Communication Power Model

Communication power Pc, in general for a bit of data, can be formulated as Pc = Phopsh=1(Pswitch(h) + Phost(h)). Where hops represents the number of hops that should be traversed between source and destination. Pswitch(h) and Phost(h) represent the power drawn in the switch/router that forwards/routes the data to the next hop and the power dissipated in the host for the same hop in case of server centric switching. In the switch centric communication within a data cen- ter, switches that connect the hosts are the major power consumption sources. In the pure server centric data center networks, servers are in charge of forwarding the data; thus, communication energy is added to the server energy profile fur- ther to the processing energy. For the hybrid network topologies, communication energy is partly dissipated in the switch and partly in the servers.

In (3.8), multiplying the PUE, data size Sdata, and replication factors r divided by throughput τc3_{, communication energy is derived from the communication power} model. ϕ(i) represents the oversubscription ratio in hop i.

Ec= (P U E × r × Sdata) h X i=1 Pswitch(i) + Phost(i) τc(i)ϕ(i) (3.8)

Moreover, the network topology impacts the power usage profile. Therefore, in the rest of this section, we discuss communication power modeling in the intra-data center communication, P2P communication and Internet communication. Data center communication power: Here, we study the power consumption of a three-tier, hierarchical topology. The motivation behind formulating the hi- erarchical model is that it can be easily generalized to numerous intra-data center topologies, e.g. Fat-Tree[5] , VL2[53], BCube[56], PCube[65], etc. The tree depth is defined based on the path messages should traverse within the data center in each layer. For the topologies which deviate from this property, e.g. CamCube[41], 3_{Note that throughput here is different than bandwidth which refers to nominal network}

Energy Analysis Metric 49 we analyze the energy model separately. We assume an l level tree in which hosts are in the leaves and are connected to an edge switch as their predecessor via Gigabit Ethernet links. The edge switches are connected via an aggregate switch; this process proceeds in two or more levels to create the root of the tree. To assign a task to a host, the root aggregate switch transmits the task data to the selected host through the tree. Assuming the homogeneous switches in each level of the tree, the power consumed for this purpose is calculated as in (3.9). Pswitchstands for the power drawn by the switch. Additionally, we added Phostto each level’s consumption to generalize our model.

Pintra−DC c = l−1 X i=1 (Pswitch(i) + Phost(i)) (3.9)

Therefore, in a switch centric model, Phost = 0, while in a pure server centric

model Pswitch = 0 and in a hybrid model, power is drawn both in switches and

servers.

Referring to (3.9), the depth of the tree, l, directly influences the power efficiency of the data center. The tree depth is determined by the number of hosts and network topology. The larger the data center is, the bigger the number of the switches and links required to connect the hosts and the deeper the tree is. Furthermore, flatter data center topologies, such as flattened butterfly [4] and FlatNet [81], obtain shorter paths via less switches. Topologies providing smaller network diameter are also more energy efficient due to shorter average path that should be traversed among the servers.

Therefore, smaller distributed data centers, serving the users independently, are more power efficient than a single mega-data center model, following a tree intra- data center topology. Loosely paraphrasing, in small data centers, the network diameter is smaller, since the number of switches and links required to connect the hosts within a data center is directly related to the number of hosts.

50 Energy Analysis Metric P2P-cloud communication power modeling: As described in the background chapter, we assume a P2P-cloud deployed in a community network. Inexpensive Wi-Fi devices have made the deployment of such communities feasible in recent years. Some flourishing instances are Guifi.net4_{, with more than 20, 000 active} nodes, Athens Wireless Metropolitan Network5_{, FunkFeuer}6_{, Freifunk}7_{, etc.} In these networks, hosts within a vicinity are usually connected via wireless links that form a wireless network. Thus, the power consumed for communication within a vicinity predominantly embraces the wireless network power consumed to trans- mit data [48].

Community networks are rather diverse in terms of size, topology and organization. This is a consequence of their unplanned deployment, based on the cooperation of their own customers/users; therefore, characterizing the power consumption in these networks is challenging. However, in the big picture, the energy consump- tion in P2P-communication platform manifests from the number of hops to be traversed to reach a particular peer, energy dissipated in each intermediate hop infrastructure such as switches, routers and antennas, and the data size Sdataand replication factor r as shown in (4.3). Dividing these terms by network through- put τh, we can calculate the energy dissipated for data transfer, since elapsed time is inversely proportional to throughput (t =_τ1

P 2P). The power consumption

is characterized in the P2P-cloud through measurement in a production wireless community network later in this chapter, in Section 3.6.

EP 2P c = X h∈hops Pc(h) ×r × Sdata τh (3.10)

Internet power consumption: P2P-clouds for inter-vicinity communication and classic data centers for communication with users rely on the Internet. Thus, to analyze the energy consumption of these systems, we should be aware of the

4_{http://guifi.net/en} 5_{http://www.awmn.net} 6_{http://www.funkfeuer.at} 7_{http://freifunk.net}

Energy Analysis Metric 51 Internet energy consumption as well. Power drawn in the Internet is subject to the hardware and distances exploited. The Internet infrastructures are classi- fied as core, distribution and access. Core layer includes the Internet backbone infrastructures such as fiber-optic channels, high speed switch/routers, etc. Distri- bution infrastructures play role as intermediaries to connect the Internet Service Providers (ISPs) to the core network. The access layer connects the user to ISP communication infrastructure.

Since there is a diverse range of hardware in each layer, it is not trivial to form a comprehensive analysis on energy consumption of the Internet. However, Baliga, et al. [10] conducted a study on the prevalent Internet hardware energy consumption. We rely on this study for the Internet power consumption part of our analysis by driving the model in (3.11). In this model, PInternet stands for the Internet power consumption which is a combination of power drawn in each level L = {core, distribution, access}. Pc(l) denotes router power consumption in layer l, and |hops(l)| indicates the number of hops, as the cardinality of the hops set in layer l, should be traversed at l layer.

PInternet=1

ϕ×

X

l∈L

Pc(l) × |hops(l)| (3.11)

In (3.12) energy consumption of communication over the Internet is modeled by dividing the power in each Internet layer by the throughput at that level τl.

EInternet= 1 ϕ× X l∈L Pc(l) × |hops(l)| τl (3.12)

The concept of oversubscription, ϕ, exists in the Internet communication, where Internet service providers exert it as a strategy to utilize the resources by over- booking the shared infrastructure among the users. The more the resources are shared temporally, the less the energy consumption is due to the shared static power dissipated. Oversubscription for the home users is 40:1 and for the business connection is around 20:1 in the current Internet.

52 Energy Analysis Metric